Long answers,
at a glance.

Type /glance after a long answer from Claude. You get it back as one plain sentence, a picture of its data and a few bullets, with every number kept.

why did our aws bill jump so much in october? cost explorer export is in billing/2026-10.csv and the terraform history is in infra/CHANGELOG.md

Why the October AWS bill jumped

The bill went from $2,110 in September to $4,760 in October (+$2,650, about 2.25×). Two Terraform changes account for about 95% of the increase. Both are inferences from lining up the cost data with the changelog dates; I haven't checked them against traffic logs.

Service Sep Oct Change
NAT Gateway $110 $2,240 +$2,130
CloudWatch $70 $450 +$380
Other $30 $125 +$95
EC2 / RDS / S3 $1,900 $1,945 +$45

1. NAT Gateway (+$2,130): ECR VPC endpoints removed on Oct 6 (#412)

  • Data going through the NAT went from 2,400 GB to 49,800 GB (about 21×).
  • At $0.045/GB, 49,800 GB comes to about $2,241, which matches the bill almost exactly. So this is processing charges, not hourly charges.
  • Removing ecr_api and ecr_dkr sends container image pulls through the NAT, where every GB is billed. The endpoints themselves only cost tens of dollars a month, so the change didn't save money overall.
  • It will be worse in November. The change was only live from Oct 6 to Oct 31, so a full month would add roughly $2,500 or more.
  • One thing to check: most of an ECR image download actually comes from S3. If the VPC has an S3 gateway endpoint (a free route to S3 that skips the NAT), most of that traffic shouldn't reach the NAT at all. 47 TB extra suggests there is no S3 gateway endpoint, or something else also started going out through the NAT. To confirm, look at Cost Explorer by day for NatGateway-Bytes (it should jump on Oct 6), or check VPC Flow Logs to see where the traffic is going.

Fix: revert #412, or at least add an S3 gateway endpoint. Also check why images are pulled so often, for example image pull policies or tasks restarting repeatedly.

2. CloudWatch (+$380): log-shipper v2.3 deployed with LOG_LEVEL=debug on Oct 3

  • 760 GB of logs went into CloudWatch, at $0.50/GB that's $380, exactly the increase.
  • Log groups are set to "Never expire", so those 760 GB, and everything logged from now on, will be stored and billed every month (about $0.03/GB-month) and that charge keeps growing.

Fix: set LOG_LEVEL back to info and give the log groups a retention period (for example 30 days).

Caveats

  • "Other" (+$95): the data doesn't show what this is. Break it down by service or usage type in Cost Explorer.
  • Not final: the CSV notes that Oct 28–31 are estimated.

Fixing both issues should bring the bill back to roughly $2,100–2,200 a month.

/glance

Your AWS bill more than doubled in October, mostly because two Terraform changes made early that month started charging for network traffic and logs.

The AWS bill rose from 2,110 dollars in September to 4,760 in October, 2.25 times, with NAT Gateway up 2,130 and CloudWatch up 380 making about 95 percent of the increase.

Of the increase, from two Terraform changes
~95%
October bill 2.25× September's

AWS bill by service

Dollars a month, largest change first. Oct 28–31 are estimated.

Service
Sep
Oct
Change
NAT Gateway2,400 → 49,800 GB processed
110
2,240
+2,130
CloudWatch760 GB logs ingested
70
450
+380
OtherNot explained yet
30
125
+95
EC2 / RDS / S3
1,900
1,945
+45
Total
2,110
4,760
+2,650

NAT Gateway: ECR endpoints removed Oct 6 (#412)

  • Without aws_vpc_endpoint.ecr_api / ecr_dkr, container image pulls go through the NAT (billed per GB of traffic): $0.045/GB ≈ $2,241. These are processing charges, not hourly ones. The endpoints only cost tens of dollars a month, so removing them saved nothing.
  • The change was only live Oct 6–31, so a full November adds ~$2,500 or more.
  • Odd: most ECR image data comes from S3, so 47 TB extra suggests there's no S3 gateway endpoint (a free route to S3), or that other new traffic is leaving through the NAT. Check: Cost Explorer by day for NatGateway-Bytes (it should jump on Oct 6), or VPC Flow Logs.

CloudWatch: debug logging since Oct 3

  • log-shipper v2.3 was deployed with LOG_LEVEL=debug, and logs are billed at $0.50/GB when they come in.
  • Log groups are set to "Never expire", so stored logs cost ~$0.03/GB-month, and that charge keeps growing.

Fixes

  • Revert #412, or at least add an S3 gateway endpoint. Check why images are pulled so often (pull policies, tasks restarting).
  • Set LOG_LEVEL back to info and add log retention (e.g. 30 days).
  • With both fixed, the bill should be ~$2,100–2,200 a month.

Still open

  • "Other" +$95: break it down by service or usage type in Cost Explorer.
  • The causes come from lining up changelog dates with the costs. They haven't been checked against traffic logs.

416 words became 1 sentence, 1 picture and 10 bullets, with every number kept.

Real answers, re-told

Every picture here is glance's own reply to a sample answer, live on the page. Sort and filter the tables, switch a panel to narrow, drag the edge to see the original, and hover any number to see the line it came from.

Jump to bills, test results, performance or progress, and see what glance does to a big table.

Bills

Why a bill jumped

Before: 276 words: the monthly totals in one line, a six-row table by service, what happened, four fixes and three caveats.

glance picked the increase as a headline figure, columns for six months of totals, then a table by service with each change in a tinted column.

Claude's answer · 276 words

why did our aws bill jump so much in october?

Why the AWS bill jumped in October

Short version: October came to $4,760, up from $2,110 in September. Almost all of the extra $2,650 is NAT Gateway and CloudWatch, and both trace back to two changes made in the first week of October.

Monthly totals

May $1,840 · June $1,910 · July $1,880 · August $2,020 · September $2,110 · October $4,760

September → October by service

Service September October Change
NAT Gateway $110 $2,240 +$2,130
CloudWatch $70 $450 +$380
EC2 $1,180 $1,210 +$30
RDS $430 $435 +$5
S3 $290 $300 +$10
Other $30 $125 +$95

What happened

  1. Debug logging left on. log-shipper v2.3, deployed Oct 3, runs with LOG_LEVEL=debug. Its output leaves the private subnets through the NAT Gateway: about 38 TB in October. The same logs also go to CloudWatch Logs (~760 GB ingested at $0.50/GB ≈ $380).
  2. ECR endpoint deleted. Terraform change #412 (Oct 6) removed the ECR VPC endpoints, so container image pulls now go through the NAT Gateway too: about 9 TB.

NAT data processed went from 2.4 TB to 49.8 TB, at $0.045/GB.

Fixes, in order

  1. Set LOG_LEVEL=info on log-shipper and redeploy. Removes ~38 TB/month of NAT traffic and most of the CloudWatch ingestion.
  2. Restore aws_vpc_endpoint.ecr_api and aws_vpc_endpoint.ecr_dkr in Terraform. Costs about $29/month (2 interface endpoints × 2 AZs) and saves roughly $400/month in NAT processing.
  3. Set CloudWatch log retention to 14 days (currently "Never expire").
  4. Add a billing alarm at $2,500.

If 1 and 2 are done by Nov 5, November should land around $2,150.

Not verified

  • Cost Explorer figures for Oct 28–31 are still estimates.
  • I didn't check Savings Plan coverage for EC2.
  • The "Other" increase (+$95) isn't explained yet.
Drag the edge to see the original. Hover a value to see where it came from.

October's AWS bill went up $2,650, almost all of it from NAT Gateway and CloudWatch, because of two changes made in the first week of October.

The AWS bill rose 2,650 dollars from September to October, mostly from NAT Gateway, whose data processed went from 2.4 to 49.8 TB.

Added from September to October
+$2,650
NAT data processed, October
49.8 TB
2.4 TB in Sep, at $0.045/GB

Monthly AWS bill

Dollars, May to October. Oct 28–31 are still estimates.

May1,840
Jun1,910
Jul1,880
Aug2,020
Sep2,110
Oct4,760

What changed, by service

Dollars, biggest change first

Service
Sep
Oct
Change
NAT Gateway
110
2,240
+2,130
CloudWatch
70
450
+380
OtherNot explained yet
30
125
+95
EC2Savings Plan coverage not checked
1,180
1,210
+30
S3
290
300
+10
RDS
430
435
+5

Teams against their budgets

Before: 152 words: a table of five teams' Q2 spend, Q3 spend and Q3 budget, why two teams went over, and what happens next.

glance picked the total as a headline figure, then a gauge per team with every budget at the same place, so the two teams over budget stand out.

Claude's answer · 152 words

how did each team do against its Q3 cloud budget?

Q3 cloud spend by team, against budget

Short version: Two of five teams went over their Q3 budget: Data platform by $8,900 and ML by $2,300. The other three came in under, but not by enough: total Q3 spend was $139,600 against $137,000 budgeted, $2,600 over.

Team Q2 Q3 Q3 budget
Data platform $41,200 $58,900 $50,000
ML $30,500 $34,300 $32,000
Web $22,400 $21,100 $25,000
Mobile backend $18,000 $17,300 $20,000
Internal tools $9,100 $8,000 $10,000

Why the two teams went over

  • Data platform: the new Spark cluster for the event pipeline has run 24/7 since July 3, on on-demand instances.
  • ML: three GPU training runs for the ranking model in August (rank-v4, rank-v4b, rank-v5).

Next

  1. Data platform moves Spark to autoscaling spot instances: estimated −$6,000 per quarter.
  2. ML asks finance to raise its Q4 budget to $36,000.

Open

  • Whether the Q4 budget for Data platform changes; finance meets on Nov 18.
Drag the edge to see the original. Hover a value to see where it came from.

Data platform and ML spent more than their Q3 budgets. That pushed total spend $2,600 over, even though the other three teams came in under.

Q3 cloud spend was $139,600 against a $137,000 budget; Data platform was $8,900 over and ML $2,300 over, while Web, Mobile backend and Internal tools stayed under.

Total Q3 spend
$139,600
Budget $137,000 · 2,600 over

Each team against its Q3 budget

Dollars. The tick on each gauge is the team's budget; red is the part past it.

Team
Q2
Q3
Budget
Data platform8,900 over
41,200
58,900
50,000
ML2,300 over
30,500
34,300
32,000
Web
22,400
21,100
25,000
Mobile backend
18,000
17,300
20,000
Internal tools
9,100
8,000
10,000

Money spent on nothing

Before: 111 words: a table of 11 idle resources in three groups with counts and monthly cost, then the status of each group.

glance picked the total as a headline figure, then a table of causes: every resource under its cause, the count, the next step and the monthly cost.

Claude's answer · 111 words

what's still running in AWS that nobody uses, and what does it cost?

What's still running that nobody uses

Short version: 11 resources cost $1,240/month and do nothing. Six of them are left over from the September load test and can be deleted today.

Cause Resources Count Monthly cost
Left over from the Sept load test lt-runner-1, lt-runner-2, lt-runner-3, lt-runner-4, lt-db, lt-cache 6 $820
Unattached EBS volumes vol-0a1f, vol-0b77, vol-0c3e 3 $210
NAT Gateways in the retired VPC nat-0f3a, nat-0f3b 2 $210

Status

  • Load test leftovers: safe to delete; the test owner confirmed on Nov 9.
  • EBS volumes: snapshot first, then delete.
  • NAT Gateways: waiting on the networking team to retire vpc-legacy.

Not checked

  • Elastic IPs and old AMIs weren't part of this sweep.
Drag the edge to see the original. Hover a value to see where it came from.

You're paying $1,240 a month for 11 AWS resources nobody uses. The six left from September's load test cost $820 of that, and you can delete them today.

Eleven unused AWS resources cost $1,240 a month: September load-test leftovers $820 (delete today), unattached EBS volumes $210 (snapshot then delete), and NAT Gateways in the retired VPC $210 (waiting on networking).

Paid for unused resources
$1,240
a month, 11 resources

What they are and what to do

Dollars a month, largest first

Cause
Count
Next step
Monthly
Sept load test leftoverslt-runner-1, lt-runner-2, lt-runner-3, lt-runner-4, lt-db, lt-cache
6
Delete today
820
Unattached EBS volumesvol-0a1f, vol-0b77, vol-0c3e
3
Snapshot, then delete
210
NAT Gateways, retired VPCnat-0f3a, nat-0f3b
2
Wait for networking
210

Test results

A nightly run in four browsers

Before: 479 words: a 24-row table of every spec in Chrome, Firefox, Safari and Edge, then each of the 5 failures and 3 flaky specs explained, and four fixes.

glance picked headline numbers and a status grid you can filter, problem specs first and the one real bug named under its spec.

Claude's answer · 479 words

the nightly e2e run is red again, what's actually failing and does any of it block thursday's release? results are in e2e/nightly-2291.txt

Nightly e2e run #2291 (main @ 4f1a9c2): what's failing

Short version: 5 of 96 runs fail, and they're all in two browsers: Safari has 3 failures and Firefox has 2. Chrome and Edge only have one flaky spec between them. Four of the five failures are test problems. One, checkout/address on Safari, is a real bug that Safari customers would hit, so I'd fix it before Thursday.

Every spec, by browser

Spec Chrome Firefox Safari Edge
auth/login pass pass pass pass
auth/signup pass pass pass pass
auth/password-reset pass pass pass pass
auth/sso-google pass pass pass pass
cart/add-item pass pass pass pass
cart/update-qty pass pass pass pass
cart/promo-code pass pass pass pass
checkout/address pass pass fail pass
checkout/shipping-rates pass pass pass pass
checkout/card-payment pass pass pass pass
checkout/apple-pay skip skip pass skip
checkout/confirmation-email pass pass pass pass
orders/history pass pass pass pass
orders/invoice-pdf pass fail pass pass
orders/return-request pass pass fail pass
account/profile pass pass pass pass
account/avatar-upload pass fail pass pass
account/addresses pass pass pass pass
search/autocomplete pass flaky pass pass
search/filters pass pass pass pass
product/gallery pass pass pass pass
product/reviews pass pass pass pass
notifications/live-updates flaky pass pass flaky
booking/date-picker pass pass fail pass

Totals: 85 pass, 5 fail, 3 flaky (passed on retry), 3 skipped (checkout/apple-pay only runs on Safari).

Safari: 3 failures

  • booking/date-picker and orders/return-request wait for #checkin and #return-date, which don't exist on WebKit: DatePicker.tsx renders a text input there instead, with the id …-text and data-testid="date-text". This is a test problem: target data-testid="date-text" on WebKit.
  • checkout/address is a real bug. AddressSuggestions.tsx:42 calls requestIdleCallback, which Safari doesn't have, so the page throws and the suggestions never appear. Safari customers can't pick an address. Fix: fall back to setTimeout when requestIdleCallback is missing.

Firefox: 2 failures

  • account/avatar-upload: AvatarButton.tsx opens the file picker from a 300 ms setTimeout, and Firefox blocks it ("picker was blocked due to lack of user activation"). Real users on Firefox hit this too, but there's a workaround (they can click again), so it can wait. For the test, use setInputFiles() on the hidden input.
  • orders/invoice-pdf waits for a download, but Firefox opens PDFs in its built-in viewer. Set pdfjs.disabled in the Firefox project's firefoxUserPrefs (playwright.config.ts doesn't set it).

Flaky (passed on retry)

  • notifications/live-updates on Chrome and Edge timed out after 2 s waiting for "Order shipped", then passed on retry.
  • search/autocomplete on Firefox waits a fixed 250 ms, but SearchBox.tsx debounces for 300 ms, so it only passes when the timing happens to work out.

What I'd do

  1. Fix requestIdleCallback in AddressSuggestions.tsx before Thursday's release.
  2. Point the two date tests at data-testid="date-text" on WebKit.
  3. Firefox: setInputFiles() for the avatar test, pdfjs.disabled for the invoice.
  4. In autocomplete, wait for the suggestions request instead of a fixed 250 ms. Give live-updates a longer timeout.

Not checked: whether checkout/address also breaks on iPad Safari. The nightly run only covers desktop WebKit.

Drag the edge to see the original. Hover a value to see where it came from.

Only one failure needs fixing before Thursday: address suggestions break for Safari customers. The other 4 failures are test problems, and nothing failed on Chrome or Edge.

Nightly run 2291: 5 of 96 runs failed, 3 on Safari and 2 on Firefox; only checkout/address on Safari is a real bug that blocks Thursday's release.

Runs failed
5
of 96; 1 blocks release
Passed
85
Flaky
3
Skipped
3

Every spec, by browser

Nightly #2291, main @ 4f1a9c2. Problem specs first; Apple Pay runs on Safari only.

7 of 24 shown
Spec
Chrome
Firefox
Safari
Edge
checkout/addressReal bug, blocks release
pass
pass
fail
pass
booking/date-picker
pass
pass
fail
pass
orders/return-request
pass
pass
fail
pass
account/avatar-upload
pass
fail
pass
pass
orders/invoice-pdf
pass
fail
pass
pass
notifications/live-updates
flaky
pass
pass
flaky
search/autocomplete
pass
flaky
pass
pass
checkout/apple-pay
skip
skip
pass
skip
auth/login
pass
pass
pass
pass
auth/signup
pass
pass
pass
pass
auth/password-reset
pass
pass
pass
pass
auth/sso-google
pass
pass
pass
pass
cart/add-item
pass
pass
pass
pass
cart/update-qty
pass
pass
pass
pass
cart/promo-code
pass
pass
pass
pass
checkout/shipping-rates
pass
pass
pass
pass
checkout/card-payment
pass
pass
pass
pass
checkout/confirmation-email
pass
pass
pass
pass
orders/history
pass
pass
pass
pass
account/profile
pass
pass
pass
pass
account/addresses
pass
pass
pass
pass
search/filters
pass
pass
pass
pass
product/gallery
pass
pass
pass
pass
product/reviews
pass
pass
pass
pass

Performance

Latency after an upgrade

Before: 367 words: the nine endpoints that miss the target listed in a sentence, how it was measured, a 40-row table of every endpoint's p95 before and after the Go 1.23 upgrade, and two caveats.

glance picked a headline count and a table you can sort, filter and search, with a gauge for each endpoint against the 300 ms target.

Claude's answer · 367 words

compare p95 for every endpoint before and after the Go 1.23 upgrade

p95 latency for all 40 endpoints after the Go 1.23 upgrade

Short version: 19 of 40 endpoints got faster, 19 got slower and 2 didn't change. 9 now miss the 300 ms p95 target: /search, /search/suggest, /products/:id, /reports/export, /notifications, /notifications/read, /addresses, /health, /feature-flags.

Measured with k6 on staging: 200 virtual users, 10 minutes per run, same data snapshot (Nov 9).

Endpoint Before (p95) After (p95) Change
/login 190 ms 180 ms −10 ms
/logout 105 ms 165 ms +60 ms
/signup 290 ms 295 ms +5 ms
/me 135 ms 95 ms −40 ms
/me/settings 80 ms 20 ms −60 ms
/search 295 ms 305 ms +10 ms
/search/suggest 225 ms 315 ms +90 ms
/products 75 ms 60 ms −15 ms
/products/:id 370 ms 380 ms +10 ms
/products/:id/reviews 270 ms 260 ms −10 ms
/cart 150 ms 290 ms +140 ms
/cart/items 105 ms 95 ms −10 ms
/checkout 175 ms 115 ms −60 ms
/checkout/confirm 205 ms 295 ms +90 ms
/orders 210 ms 195 ms −15 ms
/orders/:id 145 ms 135 ms −10 ms
/orders/:id/invoice 225 ms 260 ms +35 ms
/payments 275 ms 235 ms −40 ms
/payments/refund 255 ms 290 ms +35 ms
/shipping/rates 285 ms 295 ms +10 ms
/inventory 195 ms 170 ms −25 ms
/inventory/:sku 195 ms 200 ms +5 ms
/reports/daily 215 ms 175 ms −40 ms
/reports/export 390 ms 530 ms +140 ms
/admin/users 230 ms 170 ms −60 ms
/admin/audit 225 ms 245 ms +20 ms
/notifications 235 ms 375 ms +140 ms
/notifications/read 365 ms 350 ms −15 ms
/wishlist 300 ms 300 ms +0 ms
/wishlist/items 220 ms 220 ms +0 ms
/recommendations 325 ms 300 ms −25 ms
/coupons/apply 185 ms 175 ms −10 ms
/addresses 205 ms 345 ms +140 ms
/addresses/:id 65 ms 25 ms −40 ms
/reviews 65 ms 70 ms +5 ms
/uploads 215 ms 225 ms +10 ms
/health 380 ms 415 ms +35 ms
/feature-flags 340 ms 400 ms +60 ms
/webhooks/stripe 255 ms 230 ms −25 ms
/sitemap.xml 165 ms 125 ms −40 ms

Caveats

  • Each number is a single run; earlier run-to-run noise was about ±10 ms.
  • Staging has about 1/10 of production's data.
Drag the edge to see the original. Hover a value to see where it came from.

After the Go 1.23 upgrade, as many endpoints got slower as got faster, and 9 are now over the 300 ms target.

p95 latency for 40 endpoints before and after the Go 1.23 upgrade: 19 faster, 19 slower, 2 unchanged, and 9 now over the 300 ms target.

Now miss the 300 ms p95 target
9
of 40 endpoints
Faster
19
Slower
19
Unchanged
2

p95 latency, 40 endpoints

Milliseconds, slowest after the upgrade first. The tick on each gauge is the 300 ms target; past it is red.

12 of 40 shown
vs 300
/reports/export
390
530
+140
/health
380
415
+35
/feature-flags
340
400
+60
/products/:id
370
380
+10
/notifications
235
375
+140
/notifications/read
365
350
−15
/addresses
205
345
+140
/search/suggest
225
315
+90
/search
295
305
+10
/wishlist
300
300
+0
/recommendations
325
300
−25
/signup
290
295
+5
/checkout/confirm
205
295
+90
/shipping/rates
285
295
+10
/cart
150
290
+140
/payments/refund
255
290
+35
/products/:id/reviews
270
260
−10
/orders/:id/invoice
225
260
+35
/admin/audit
225
245
+20
/payments
275
235
−40
/webhooks/stripe
255
230
−25
/uploads
215
225
+10
/wishlist/items
220
220
+0
/inventory/:sku
195
200
+5
/orders
210
195
−15
/login
190
180
−10
/reports/daily
215
175
−40
/coupons/apply
185
175
−10
/inventory
195
170
−25
/admin/users
230
170
−60
/logout
105
165
+60
/orders/:id
145
135
−10
/sitemap.xml
165
125
−40
/checkout
175
115
−60
/me
135
95
−40
/cart/items
105
95
−10
/reviews
65
70
+5
/products
75
60
−15
/addresses/:id
65
25
−40
/me/settings
80
20
−60

What a bundle is made of

Before: 183 words: what's in main.js as a six-row table, three fixes with their savings, and three things to watch out for.

glance picked a split bar and a table of what it's made of, then the fixes as bars, biggest win first.

Claude's answer · 183 words

why is main.js so big now?

Why the dashboard bundle is 1.84 MB

Short version: Almost half of the main bundle is the charting library, and it loads on every page even though only the Reports page uses it. Loading it only there would cut main.js by about 860 KB.

What's in main.js (1.84 MB minified, 512 KB gzipped)

Part Size (minified) Share
echarts 860 KB 47 %
Our code 410 KB 22 %
react + react-dom 140 KB 8 %
lodash 72 KB 4 %
moment + locales 290 KB 16 %
Everything else (31 packages) 68 KB 4 %

From npx vite-bundle-visualizer on commit a41c9e2.

What to do, biggest win first

  1. Load echarts with React.lazy in pages/Reports.tsx only: −860 KB from main.js (it moves to a reports chunk).
  2. Replace moment with date-fns and drop the unused locales: about −250 KB.
  3. Change import _ from 'lodash' to per-function imports (lodash/debounce): about −60 KB.

Watch out

  • The Reports page will be about 300 ms slower on a first visit, while the chart chunk downloads.
  • react-datepicker also uses moment; check it still works after step 2.
  • I haven't measured the gzipped size after these changes.
Drag the edge to see the original. Hover a value to see where it came from.

main.js is big mostly because the chart library, echarts, loads on every page even though only the Reports page uses it. It makes up almost half the file.

main.js is 1.84 MB minified; echarts is 860 KB of it (47 percent), and three fixes would cut about 860, 250 and 60 KB.

What's in main.js

1.84 MB minified, 512 KB gzipped. Sizes in KB minified, left to right.

From npx vite-bundle-visualizer on commit a41c9e2

Part, left to right
Size (KB)
Share (%)
echartsOnly Reports uses it
860
47
Our code
410
22
moment + locales
290
16
react + react-dom
140
8
lodash
72
4
Everything else31 packages
68
4

Fixes, biggest win first

KB cut from main.js

Lazy-load echarts
860
moment to date-fns
~250
Per-function lodash
~60

Progress

A day of work, recapped

Before: 526 words over 4 answers: four answers in one session: a first measurement against the budget, two rounds of fixes with new numbers, then what's left and the plan for Monday.

Recapped with /glance everything we did today. glance picked one row per metric with the start, step 1 and now, and a gauge against its budget.

Claude's answer · 526 words

checkout feels slow on mobile. can you measure it against our perf budget? lighthouse results are in perf/lighthouse-checkout.json

Checkout on mobile, against the performance budget

Short version: All five metrics are over budget. The page takes 4.8 s to show its main image, and the three biggest problems are an unoptimized hero image, a banner that pushes the page down after it loads, and 612 KB of JavaScript.

Median of 5 Lighthouse runs on a Moto G Power profile (slow CPU, Slow 4G):

Metric Now Budget
LCP (main content visible) 4.8 s 2.5 s over
INP (response to a tap) 380 ms 200 ms over
CLS (layout shift) 0.21 0.1 over
JavaScript, gzipped 612 KB 300 KB over
TTFB (server response) 0.9 s 0.8 s over

Why

  • LCP: the hero is checkout-hero.png, a 1.4 MB PNG that isn't preloaded, so it starts downloading late.
  • CLS: .promo-banner is inserted about 1.9 s after load and pushes everything down (0.19 of the 0.21).
  • JavaScript: the Places SDK (180 KB) and the payments SDK (210 KB) load on every visit, before anyone reaches those steps. moment adds 72 KB to main.js.
  • INP: every keystroke in the coupon field runs CouponField.onChange, a 140 ms task.
  • TTFB: the page waits on GET /api/cart (610 ms), which is sent with no-store, so it's never cached.

I'd start with the image and the banner: both are small changes and should fix two metrics.

converted the hero to avif with a preload and reserved space for the banner. new run is in perf/run-2.json

Both changes worked. Median of 5 runs, same device profile:

Metric Before After Budget
LCP 4.8 s 2.9 s 2.5 s
CLS 0.21 0.04 0.1
INP 380 ms 370 ms 200 ms
  • CLS is now well under budget. The reserved slot means nothing moves when the banner arrives.
  • LCP is better but still 0.4 s over. The image itself now loads in 310 ms, but the browser waits 1,570 ms after that before painting it, because it's busy running JavaScript. That's the next fix: the 612 KB of scripts.
  • INP and JavaScript didn't change, as expected.

lazy loaded the places and payments sdks and swapped moment for Intl. perf/run-3.json

That did it for JavaScript and LCP. Median of 5 runs:

Metric Start of today Now Budget
JavaScript, gzipped 612 KB 284 KB 300 KB
LCP 4.8 s 2.3 s 2.5 s
INP 380 ms 240 ms 200 ms
  • JavaScript and LCP are under budget. With less script to run, the hero paints sooner.
  • INP improved but is still 40 ms over. The main thread is less busy, but CouponField.onChange still runs a 140 ms task on every keystroke.

what's left? we'll ship what we have behind the checkout-v2 flag on monday and do INP next week

What's left

  • INP, 240 ms (budget 200 ms): the coupon field still validates on every keystroke. Debouncing it (250 ms) and moving the validation into a web worker should bring it under 200 ms, but that's an estimate until it's measured.
  • TTFB, 0.9 s (budget 0.8 s): unchanged all day. None of today's changes touched the server; the /api/cart response would need caching on the backend.

For Monday

  • Roll out behind checkout-v2 at 10% of traffic first, and go to 100% only if the checkout error rate stays flat.
  • Add a Lighthouse CI check with the budget so LCP, CLS and JavaScript can't creep back over. INP can't be measured in a lab run, so use Total Blocking Time (200 ms) as a stand-in.
  • INP is next week. TTFB needs the backend team.
Drag the edge to see the original. Hover a value to see where it came from.

Checkout on mobile now loads fast enough and stays put on screen. Taps and the server are still slower than the budget allows.

Checkout mobile metrics at the start, after step 1 and now, against budget: LCP 4.8, 2.9, 2.3 s (budget 2.5); INP 380, 370, 240 ms (budget 200, still over); CLS 0.21, 0.04, 0.04 (budget 0.1); JavaScript 612, 612, 284 KB (budget 300); TTFB 0.9 s throughout (budget 0.8, still over).

Checkout on mobile, against the budget

Median of 5 Lighthouse runs, Moto G Power, Slow 4G. The tick on each gauge is the budget.

Step 1: hero image and banner slot. Step 2: lazy SDKs, Intl for moment.

Metric
Start
Step 1
Budget
Now
LCP (s)Main content visible
4.8
2.9
2.5
2.3
INP (ms)40 over
380
370
200
240
CLSLayout shift
0.21
0.04
0.1
0.04
JavaScript (KB gzip)
612
612
300
284
TTFB (s)Server response
0.9
0.9
0.8
0.9

Big tables

A wide table is where a plain answer reads worst. The chat gives every column with a long cell the room, so short columns shrink to stacks of wrapped words and rows grow four lines tall. glance rebuilds the same table: the same 14 rows and numbers, one line each, grouped by what to do and sorted by what you'd save.

Stock Claude · 622 words, the table as the chat lays it out

Our SaaS renewals for early 2027 are in finance/saas-renewals-2027.csv. Which ones should we cut or shrink before they renew, and how much would we save? Give me the full picture for every tool.

The clearest savings are about $46.6k a year from six tools: Figma, Notion, Zoom, Miro, Linear and DocuSign. Removing unused seats on five more tools adds about $3.8k, for about $50k in total. That's roughly 24% of the $205k annual spend. Several of these renew in January, so timing matters (see the end).

Tool Renews Cost Seats used (90 days) / paid Recommendation Annual saving
Figma Jan 15 $21,600 18 / 40 Shrink to about 18 seats ($540 each) $11,880
Miro Jan 20 $6,300 9 / 35 Cut. FigJam, which comes with Figma, covers most of what it's used for $6,300
DocuSign Jan 25 $4,800 3 / 10 Shrink to 3 seats, or switch to a per-envelope plan (about 60 envelopes a year) $3,360 at least, more if you switch plans
Linear Jan 31 $4,800 6 / 25 Cut. The pilot ended and the team went back to Jira $4,800
Notion Feb 1 $17,280 64 / 120 Shrink to about 64 seats ($144 each), then decide whether to merge it into Confluence $8,064
Calendly Feb 14 $3,840 19 / 20 Keep. Drop 1 seat $192
Zoom Feb 20 $16,200 22 / 90 Shrink to about 22 seats ($180 each), or cut it if those users can move to Meet $12,240 (up to $16,200)
Slack Mar 1 $25,200 131 / 140 Keep. Drop 9 seats about $1,750 against the new price
Datadog Mar 31 $48,000 committed n/a Raise the commitment (see below) Avoids paying for overage
Confluence Apr 10 $6,480 57 / 60 Keep. It comes bundled with Jira –
Jira Apr 10 $9,360 58 / 60 Keep –
Loom May 5 $4,500 27 / 30 Keep. Drop 3 seats $450
Google Workspace Jun 30 $23,520 138 / 140 Keep. Drop 2 seats $336
1Password Aug 1 $13,440 129 / 140 Keep, since security policy requires it. Drop 11 seats $1,056

Then three more sections: what changes the numbers, Datadog, and timing.

glance · 219 words around one picture

Cutting Miro and Linear and shrinking four other tools saves about $46.6k a year. Removing spare seats elsewhere brings the total to about $50k.

About $50.4k a year can be saved across 14 SaaS renewals: cut Miro and Linear, shrink Zoom, Figma, Notion and DocuSign, trim seats on five tools, keep Jira and Confluence, and raise the Datadog commitment.

Saving a year
~$50.4k
~24% of $205k spend
Cut or shrink 6 tools
~$46.6k
Drop unused seats on 5 tools
~$3.8k

Every tool

Dollars a year, biggest saving first in each group

Grey is seats used in the last 90 days, blue is unused seats

Tool
Renews
Used / paid
Cost
Saving
Cut
MiroFigJam covers most use
Jan 20
9 / 35
6,300
6,300
LinearPilot ended; back to Jira
Jan 31
6 / 25
4,800
4,800
Shrink
Zoom~22 seats; 16,200 if cut
Feb 20
22 / 90
16,200
12,240
Figma~18 seats at $540
Jan 15
18 / 40
21,600
11,880
Notion~64 seats at $144
Feb 1
64 / 120
17,280
8,064
DocuSign3 seats or per-envelope
Jan 25
3 / 10
4,800
>3,360
Keep, drop seats
SlackDrop 9; price +8%
Mar 1
131 / 140
25,200
~1,750
1PasswordDrop 11; policy requires it
Aug 1
129 / 140
13,440
1,056
LoomDrop 3
May 5
27 / 30
4,500
450
Google WorkspaceDrop 2
Jun 30
138 / 140
23,520
336
CalendlyDrop 1
Feb 14
19 / 20
3,840
192
Keep
Jira
Apr 10
58 / 60
9,360
—
ConfluenceBundled with Jira
Apr 10
57 / 60
6,480
—
Raise the commitment
Datadog48,000 committed, ~61,000 used
Mar 31
n/a
48,000
—

A renewal review for a made-up 140-person company, answered by stock Claude (Opus 5.5), then re-told with /glance. Both columns are the same width.

Install in one line

git clone https://github.com/haslanli/glance-skill ~/.claude/skills/glance

Claude Code picks up new skills while it runs, so there's no restart. If ~/.claude/skills didn't exist before, run /reload-skills once. To update later, run git -C ~/.claude/skills/glance pull.

Commands

After a long answer, type /glance in the main conversation.

/btw /glance doesn't work as well: side questions can't load skills, so Claude guesses at the format from its one-paragraph description. In our tests it kept 52–63% of the numbers and names, against 99% for /glance.

CommandWhat it does
/glanceRe-tells the last answer once
/glance <what>Re-tells one answer, part of one, or several: your reply about the build, only the numbers, everything we did today, all decisions so far, or pasted text
/glance onRe-tells the last answer, then answers this way until you turn it off
/glance superSame as on, with every bullet cut to a short fragment
/glance offBack to normal replies
/glance statusSays which mode is on
/glance helpShows this list

Why not just use dataviz?

dataviz draws the chart you ask for. glance re-tells an answer you already have.

  • It decides what's worth a picture. Whether anything in the answer is, which part, and which form shows it best. Then it rewrites the words around it.
  • It keeps every fact and invents nothing. Every number, name, path, command, warning and open question stays in. No trend lines, averages or projections; a missing value shows as —.
  • The answer comes first. One plain sentence, then the picture, then a few bullets. You have the point before you read a single axis.
  • It only draws when there's data. Reasoning stays as words, so you never get a chart for the sake of a chart.
  • Every picture comes from one tested kit, checked in light mode, dark mode and narrow side panels. Its thin marks, one-axis rule and color checks come from dataviz's own rules.

Use dataviz when you want Claude to make a chart. Use glance when Claude has already answered and you want to understand it fast.

Readable by everyone

About one man in twelve can't easily tell red from green. Most charts lean on exactly that pair, so glance is built to work without it.

  • Blue says look here, red says negative. Red appears only on a value past a limit, target or budget, or on something failing, and never alone: the value also runs past the tick on its gauge, or says "fail".
  • No green, no amber. An improvement stays soft blue. Amber and the alert red differ by only 3.2 for red-green color-blind readers, where 8 is the minimum.
  • Every value is printed, and each picture opens with a one-sentence summary for screen readers.

Q3 cloud spend was $139,600 against a $137,000 budget; Data platform was $8,900 over and ML $2,300 over, while Web, Mobile backend and Internal tools stayed under.

Total Q3 spend
$139,600
Budget $137,000 · 2,600 over

Each team against its Q3 budget

Dollars. The tick on each gauge is the team's budget; red is the part past it.

Team
Q2
Q3
Budget
Data platform8,900 over
41,200
58,900
50,000
ML2,300 over
30,500
34,300
32,000
Web
22,400
21,100
25,000
Mobile backend
18,000
17,300
20,000
Internal tools
9,100
8,000
10,000
Red and amber3.2too close
Red and green2.1too close
Red and blue22.1clear

Difference under deuteranopia, OKLab ΔE × 100, from the dataviz skill's palette validator. The toggle simulates it with the Machado, Oliveira and Fernandes (2009) matrix.

Credits

glance started from tldr by Generative Labs. Pointing at an earlier answer, the on, super and off modes, and the rule to cut wording but keep every fact all come from tldr. glance adds the plain-language first line, the pictures and the kit. If you only want bullet summaries, tldr does that well.

What a bill answer needs to show adapts ideas from Anthropic's variance-analysis skill, AWS's billing and cost management skill and the cost-check skill from aws-samples. The examples on this page are glance's replies to sample answers written for its tests.